Electric power facility intelligent anti-climbing early warning method, system, equipment and medium

Abstract

The invention discloses an intelligent anti-climbing early warning method, system, equipment and medium for electric power facilities, and belongs to the technical field of telegraph pole tower monitoring, and the method comprises the steps: an image collection module captures an image sequence of a telegraph pole tower area through a fixed camera, a preprocessing module carries out the preprocessing through employing OpenCV according to the sequence of denoising, graying, morphological operation and a frame difference method, and a pre-processing module carries out the pre-processing according to the sequence of the denoising, graying, morphological operation and a frame difference method; the feature extraction module adopts a multi-stage neural network architecture to output an optimized spatial-temporal feature matrix; the behavior recognition module adopts a CNN + Softmax architecture to judge whether a climbing behavior exists or not; the time sequence analysis module is used for carrying out time sequence analysis on the identified climbing behaviors by adopting a long short-term memory network; and the early warning and response module implements a dynamic threshold strategy to perform behavior confirmation and trend verification. According to the method, the monitoring accuracy is remarkably improved, the monitoring cost is reduced, the algorithm adaptability is enhanced, and real-time early warning is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of electric pole tower monitoring, and specifically relates to an intelligent anti-climbing early warning method, system, device and medium for power facilities. BACKGROUND

[0002] Currently, electric pole tower monitoring technology mainly includes video monitoring and sensor monitoring. Video monitoring technology captures images of the tower area through the installation of cameras and analyzes the images through manual or automated algorithms to detect abnormal behavior. Sensor monitoring technology monitors the state of the tower and changes in the surrounding environment in real time by installing various sensors (such as vibration sensors, temperature sensors, etc.) on the tower.

[0003] The existing technology has the following disadvantages: In video monitoring technology, it is susceptible to environmental interference. In outdoor environments, factors such as changes in light, wind and rain, and obstructions can affect the image quality captured by the camera, thereby affecting the monitoring effect. Manual monitoring is costly, and long-term manual monitoring not only increases labor costs, but also is prone to false positives or false negatives due to fatigue and other reasons. Algorithm limitations: Traditional image analysis algorithms have limited recognition accuracy in complex scenarios, making it difficult to effectively identify complex behaviors such as climbing.

[0004] In sensor monitoring technology, the arrangement of sensors is complex, multiple sensors need to be installed on the tower, the wiring is complex, and the maintenance cost is high. The monitoring range is limited, and sensors can usually only monitor specific physical quantities, with limited direct monitoring capabilities for climbing and other behaviors. Susceptible to interference, sensors are susceptible to environmental noise and electromagnetic interference, resulting in inaccurate data. SUMMARY

[0005] In view of the above existing problems, the present application proposes an electric pole tower climbing behavior monitoring system based on OpenCV, neural networks and advanced graph convolution technology. The system aims to integrate image processing and deep learning technology, especially the spatio-temporal graph convolution network (ST-GCN), the adaptive graph convolution network (AGCN), the spatio-temporal attention module (AST), and the skeleton joint enhanced traversal space configuration strategy, to realize real-time monitoring of the electric pole tower area, accurately identify climbing behavior, and timely trigger an early warning signal, thereby effectively ensuring the safe operation of the power system.

[0006] The technical problems to be solved by the present application include: significantly improving monitoring accuracy, using advanced graph convolution technology and deep learning models to improve the recognition accuracy of climbing behavior; reducing monitoring costs: without the need to install a large number of sensors on the tower, only using existing camera equipment, realizing monitoring functions through software algorithms, and reducing hardware costs; enhancing algorithm adaptability: designing a robust neural network model to improve the resistance of the algorithm to different interference factors such as light conditions, occlusion, and noise; realizing real-time early warning: constructing an early warning and response mechanism, triggering an early warning signal as soon as climbing behavior is detected, and formulating a corresponding response strategy.

[0007] To solve the above technical problems, an intelligent anti-climbing early warning method for power facilities is proposed, comprising,

[0008] The image acquisition module captures image sequences of the power line tower area through a fixed camera and outputs the original image sequences to the preprocessing module; the preprocessing module uses OpenCV to preprocess according to the order of denoising, graying, morphological operation and frame difference method, and outputs the optimized gray image and motion area mask to the feature extraction module; the feature extraction module uses a multi-level neural network architecture to construct a spatio-temporal graph, dynamically adjust the graph structure, weight the spatio-temporal attention, and allocate dynamic weights, and outputs an optimized spatio-temporal feature matrix; the behavior recognition module uses a CNN+Softmax architecture to extract local features through a convolutional neural network, calculates the behavior probability through Softmax, and outputs the climbing behavior probability value to determine whether there is climbing behavior; the time series analysis module uses a long short-term memory network to analyze the time series of the identified climbing behavior and predicts the duration and trend of the climbing behavior; the early warning and response module implements a dynamic threshold strategy for behavior confirmation and trend verification, triggers a response strategy, obtains a hierarchical early warning signal and a response instruction, and implements global optimization through the system integration and optimization module; wherein the fixed camera is selected to have high resolution, wide dynamic range, night vision performance and protection characteristics; the original image sequence includes light changes and moving targets.

[0009] As a preferred scheme of the intelligent anti-climbing early warning method for power facilities, the preprocessing includes using the OpenCV library to preprocess the collected images according to the order of denoising, graying, morphological operation and frame difference method.

[0010] The denoising performs smoothing processing on the video image, and performs denoising processing on the image through Gaussian filtering.

[0011] The graying converts a color image into a gray image, and only retains the brightness information of the image.

[0012] The morphological operation removes noise points by performing an expansion operation that expands white regions in a binary image, a corrosion operation that reduces white regions, and filling holes in the image, and cleans up irregular shapes on the image boundary through the morphological operation.

[0013] The image processed through noise removal, grayscale, and morphological operation is processed by frame difference method, and a difference image is obtained by subtracting the current frame from the previous frame, threshold processing is applied to highlight the change area, and the image after grayscale processing is combined to locate the human posture.

[0014] As a preferred scheme of the power facility intelligent anti-climbing early warning method, the multi-level neural network architecture includes a space-time graph convolution network for space-time graph construction and an adaptive graph convolution network for dynamic adjustment of the graph structure.

[0015] The space-time graph convolution network converts the preprocessed image into skeleton joint information, constructs a space-time graph, extracts space-time features through multiple graph convolution layers, and constructs a feature matrix, and each layer of graph convolution transmits information between nodes through an adjacency matrix.

[0016] The adaptive graph convolution network inputs the feature matrix output by the space-time graph convolution network, which contains the space-time features extracted by the space-time graph convolution network.

[0017] The adaptive graph convolution network introduces an adaptive adjacency matrix The graph convolution network dynamically adjusts the edge weight according to the relationship between the actual data, dynamically adjusts the graph structure, and the update mode of the adaptive adjacency matrix is as follows:

[0018]

[0019] Wherein, Z (l+1) is the feature matrix of the graph convolution in the space-time graph convolution network, and σ1 is a nonlinear activation function ReLU, is an adjacency matrix with a self-loop, A is an original adjacency matrix representing the connection relationship between the joints; I N is an identity matrix for introducing a self-loop connection; Y (l+1) is the updated feature matrix, which dynamically adjusts the edge weight of the graph according to the learned parameters; W (l) is the weight matrix of the lth layer, is a new degree matrix calculated based on the adaptive adjacency matrix.

[0020] As a preferred scheme of the power facility intelligent anti-climbing early warning method, the multi-level neural network architecture further includes a space-time attention module for space-time attention weighting and a skeleton joint node enhanced space configuration strategy for dynamic weight distribution.

[0021] The spatio-temporal attention module performs attention weighting on the updated feature matrix, calculates a spatial attention score to show the importance of different joint nodes, and obtains key joint nodes according to the importance of different joint nodes, so as to highlight the influence of the key joint nodes on behavior recognition;

[0022] The spatial attention score is used to weight the updated feature matrix to generate a spatial weighted feature matrix;

[0023] The attention mechanism is used to calculate the relevance between different time points. For each time step, the non-normalized score of the current time step is calculated and normalized to a time attention score by a Softmax function;

[0024] The time attention score is used to weight the spatial weighted feature matrix to generate a time weighted feature matrix;

[0025] The time weighted feature matrix and the updated feature matrix are fused to generate the final weighted feature matrix of the spatio-temporal attention module

[0026]

[0027] Wherein, MLP is a multi-layer perception machine used to fuse attention weighted features; + represents a residual connection to retain original feature information; Y time is a time weighted feature matrix;

[0028] The bone joint reinforced traversal space configuration strategy is used to fuse the final weighted feature matrix output by the spatio-temporal attention module The dimension NxF is set, where N is the number of joint nodes and F is the feature dimension;

[0029] The spatial configuration optimization is performed to decode the spatial coordinates of the joint nodes from the feature matrix, input the final weighted feature matrix to calculate the importance weight of the dynamic weight between each pair of joint nodes by a lightweight network and generate an importance weight matrix W, calculate the weighted spatial distance according to the importance weight and generate a weighted spatial distance matrix;

[0030] The hierarchical traversal is performed on the weighted spatial distance, which is divided into local traversal and global traversal. The local traversal is to preferentially traverse the connection between the joint nodes of the limbs to capture local motion details. The global traversal is to aggregate local features based on dynamic weights to generate a global motion pattern description;

[0031] In the hierarchical traversal process, the spatial relationship between the joint nodes is optimized by the importance weight and the weighted spatial distance, and the traversal result is combined with the final weighted feature matrix to generate an optimized joint feature matrix F ' , which is calculated as follows:

[0032]

[0033] Wherein, WM is a dynamic weight matrix, Dis is a weighted spatial distance matrix, MLP is a multi-layer perceptron, and is used to fuse spatial optimization information.

[0034] As a preferred scheme of the power facility intelligent anti-climbing early warning method, the judgment of whether there is a climbing behavior includes using a CNN+Softmax architecture including a convolutional neural network and a Softmax behavior classification.

[0035] The convolutional neural network includes inputting the optimized joint feature matrix into the convolutional neural network of the behavior recognition module, including a convolutional layer, a pooling layer and a fully connected layer.

[0036] The convolutional layer extracts local features of the human body posture.

[0037] The pooling layer performs down-sampling on the extracted local features, flattens the pooled features into a one-dimensional vector, and inputs them into the fully connected layer.

[0038] The fully connected layer maps the features processed by the pooling layer to a classification space to determine whether there is a climbing behavior.

[0039] The Softmax behavior classification includes using a Softmax function to calculate the probability distribution of each class to determine whether it belongs to a climbing behavior.

[0040]

[0041] Wherein, P(y=i|x) is the probability that the input x belongs to class i, z i is the output value of the i-th neuron, i.e. the result of weighted summation; M is the total number of classes, and j is a variable index.

[0042] As a preferred scheme of the power facility intelligent anti-climbing early warning method, the time series analysis includes using a long short-term memory network to perform time series analysis on the identified climbing behavior, and predicting the duration and trend of the climbing behavior by constructing a long short-term memory network model.

[0043] f t =σ2(W f ·[h t-1 ,x t ]+b f )

[0044] i t =σ2(W i ·[h t-1 ,x t ]+bi )

[0045]

[0046] o t = sigma2(W o * [h t-1 , x t ] + b o )

[0047] h t = o t * tanh(C t )

[0048] Wherein, f t , i t , o t It is the output value of the forget gate, input gate and output gate respectively, C t , It is the cell state and candidate cell state respectively, h t It is the hidden layer output value, W f , W i , W C , W o It is the weight matrix of the forget gate, input gate, cell state and output gate respectively, b f , b i , b C , b o It is the corresponding bias term respectively, sigma2 is the sigmoid activation function, and tanh is the hyperbolic tangent activation function.

[0049] As a preferred scheme of the power facility intelligent anti-climbing early warning method, wherein: the behavior confirmation and trend verification includes a detection stage, a time series analysis stage and a warning stage;

[0050] The detection stage judges whether the climbing behavior occurs, compares the probability value P(y=1∣x) output by the behavior recognition module with the detection threshold T detect , and triggers different levels of warning signals:

[0051] When P(y=1∣x)≥T detect , it is determined that the climbing behavior is confirmed, a red emergency warning is carried out, and a warning signal is sent out through sound and light alarm and platform pop-up window, the alarm is automatically pushed to the monitoring center, the unmanned aerial vehicle or security personnel on-site verification is started, and the video evidence is recorded;

[0052] When P(y=1∣x)<T detect , but P(y=1∣x)≥0.5, it is determined that it is a suspected climbing behavior, which needs to be further verified in combination with the time series analysis stage;

[0053] The time series analysis stage predicts whether the climbing behavior is continuous, analyzes the behavior probability trend of continuous frames through the time series module, inputs the behavior probability sequence {P t ,P t-1 ,...,P t-K} of the past K frames, predicts the continuous probability of the behavior in the future Δt seconds through the long short-term memory network of the time series module, and if the continuous probability reaches the probability threshold P continue , it is determined that the climbing behavior will continue.

[0054] When the suspected climbing behavior and the climbing behavior are determined to continue, a yellow boundary warning is carried out, and a warning signal is sent out through a platform pop-up window and a short message notification, a video stream is manually reviewed, and an emergency response is prepared;

[0055] When P(y=1∣x)<T detect , P(y=1∣x)<0.5 but there is abnormal motion, a blue observation warning is carried out, and no warning signal is sent out through platform registration, the system automatically tracks the target, does not trigger active alarm, and only records logs.

[0056] Another object of the present application is to provide an intelligent anti-climbing warning system for power facilities.

[0057] To solve the above technical problems, the present application provides the following technical scheme: an intelligent anti-climbing warning system for power facilities, comprising an image acquisition module, a preprocessing module, a feature extraction module, a behavior recognition module, a time series analysis module, a warning and response module, and a system integration and optimization module.

[0058] The image acquisition module uses a fixed camera to capture image sequences of the power pole tower area as the basis for subsequent processing.

[0059] The preprocessing module uses the OpenCV library to preprocess the collected images according to the order of denoising, grayscale, morphological operation and frame difference method.

[0060] The feature extraction module is responsible for extracting and identifying key features of climbing behavior from the preprocessed images, including spatio-temporal graph convolution network, adaptive graph convolution network, spatio-temporal attention module and skeleton joint enhanced traversal space configuration strategy.

[0061] The behavior recognition module uses a trained neural network model to perform deep analysis on the features output by the feature extraction module to determine whether there is a climbing behavior.

[0062] The time series analysis module uses a long short-term memory network to perform time series analysis and prediction on the identified climbing behavior in the time series analysis stage.

[0063] The pre-warning and response module triggers a pre-warning signal and formulates a corresponding response strategy when detecting a climbing behavior or predicting that the climbing behavior will continue to develop according to the prediction result of the time series analysis module.

[0064] The system integration and optimization module integrates all the modules to form a monitoring system, and optimizes the algorithm.

[0065] The application provides a computer device, including a memory and a processor, the memory stores a computer program, characterized in that the processor implements the steps of the power facility intelligent anti-climbing pre-warning method when executing the computer program.

[0066] The application provides a computer readable storage medium, which stores a computer program, characterized in that the computer program is executed by a processor to implement the steps of the power facility intelligent anti-climbing pre-warning method.

[0067] The application has the following beneficial effects:

[0068] 1. Significantly improve monitoring accuracy: By fusing OpenCV image preprocessing algorithm and advanced graph convolution technology, as well as deep learning feature extraction and behavior recognition algorithm, the recognition accuracy of climbing behavior is significantly improved.

[0069] 2. Reduce monitoring cost: Without installing a large number of sensors on the tower, only the existing camera equipment can realize the monitoring function, reducing the hardware cost.

[0070] 3. Enhance algorithm adaptability: The designed neural network model has strong robustness and can cope with different interference factors such as light conditions, occlusion and noise.

[0071] 4. Realize real-time pre-warning: The pre-warning and response mechanism is constructed, once the climbing behavior is detected or the climbing behavior is predicted to continue to develop, the pre-warning signal can be triggered immediately and the corresponding response strategy can be formulated. BRIEF DESCRIPTION OF DRAWINGS

[0072] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings:

[0073] Figure 1 The overall flowchart of the power facility intelligent anti-climbing pre-warning method provided by an embodiment of the application.

[0074] Figure 2The scheme module diagram of the power facility intelligent anti-climbing early warning system provided for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0075] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0076] Embodiment 1, refer to Figure 1 For the first embodiment of the present application, the embodiment provides a power facility intelligent anti-climbing early warning method, comprising:

[0077] S1: The image acquisition module captures the image sequence of the power line tower area through the fixed camera, and outputs the original image sequence to the preprocessing module; the original image sequence includes illumination variation and moving target.

[0078] It should be noted that the fixed camera is used to capture the image sequence of the power line tower area as the basis for subsequent processing.

[0079] In the selection of the camera, consider the model with high resolution, wide dynamic range and good night vision performance to ensure that clear and detailed images can be captured under different lighting conditions.

[0080] At the same time, the camera must have dustproof, waterproof, shockproof and other characteristics to adapt to outdoor harsh environments. The selection of the camera is crucial because it directly determines the quality of the input data.

[0081] The ideal camera should be able to provide stable and high-quality images under various lighting conditions.

[0082] Taking this into account, the camera with the following characteristics should be selected:

[0083] High resolution: ensures that the image contains enough detailed information.

[0084] Wide dynamic range: can maintain good image quality in extreme brightness differences.

[0085] Night vision performance: can work normally in low light environment.

[0086] Protective features: dustproof, waterproof, shockproof, etc., to ensure long-term stable operation of the device.

[0087] These characteristics make the camera not only can obtain high-quality images in the day, but also can maintain high imaging quality in the night or bad weather conditions, thereby providing a solid data foundation for subsequent image processing.

[0088] S2: The preprocessing module uses OpenCV to perform preprocessing according to the order of denoising, graying, morphological operation and frame difference method, and outputs the optimized gray image and motion area mask to the feature extraction module.

[0089] It should be noted that the collected image is preprocessed according to the order of denoising, graying, morphological operation and frame difference method using the OpenCV library; wherein the first step is to remove noise, otherwise the subsequent operation may amplify the noise value. The second step of graying can reduce the complexity of color images, and morphological operation is usually applied to gray images. After the above operations, the frame difference method is finally processed to improve the motion detection capability.

[0090] The denoising, the video image is smoothed, and the image is denoised by Gaussian filtering;

[0091] Specifically, since the image is often disturbed by various noises such as device noise and environmental noise during the collection process, these noises will seriously affect the quality of the image, and then affect the subsequent feature extraction and behavior recognition effect. The video image is smoothed to effectively filter out the noise in the video image. For this purpose, the image is denoised by using Gaussian filtering. Gaussian filtering is a linear smoothing filter that removes high-frequency components by weighted averaging and retains low-frequency components to achieve the purpose of noise reduction. Algorithm implementation:

[0092]

[0093] Wherein, G(alpha, beta) is a Gaussian function, (alpha, beta) is a pixel coordinate, and mu is the standard deviation of the Gaussian filter. The denoised image reduces the interference of irrelevant information, so that the subsequent processing can focus more on important visual features. This not only improves the effect of graying, but also lays a better foundation for morphological operation.

[0094] The graying converts the color image into a gray image, only retaining the brightness information of the image; thereby simplifying the subsequent processing procedure. Graying processing not only reduces the data volume, but also improves the processing speed. Through the above series of preprocessing steps, this module effectively improves the quality of the image, and provides more clear and accurate image data for subsequent feature extraction and behavior recognition:

[0095] Gray = 0.299 * R + 0.587 * G + 0.114 * B

[0096] Wherein, Gray is the gray value, R, G, B are the pixel values of red, green and blue channels respectively. This formula reflects the different sensitivities of the human eye to different colors, with the green channel having the largest weight, the red channel second, and the blue channel the smallest. After gray scaling, the image only retains the brightness information, eliminating the complexity brought by color changes. This is particularly advantageous for subsequent morphological operations, as the latter relies on binary image processing, which is easier to perform on a grayscale image. In addition, gray scaling reduces the amount of calculation, improves processing speed, and also reduces storage requirements.

[0097] The morphological operation removes small noise points and fills the holes in the image by performing dilation and erosion operations. The dilation operation expands the white area in the binary image, and the erosion operation reduces the white area. The morphological operation cleans up the irregular shapes on the image boundary; for the dilation operation, assuming the structure element is Q, the dilated image E can be represented by the following formula: And for the erosion operation, there is E represents the original binary image, Q is the structure element, and Θ represent the dilation and erosion operators respectively, and z represents the center point of the structure element Q on the image E.

[0098] The image processed by noise removal, gray scaling and morphological operation is processed by frame difference method, and the difference image is obtained by subtracting the current frame from the previous frame. Threshold processing is applied to highlight the change area, and the human body posture is located in combination with the gray scaled image.

[0099] Specifically, let the current frame be I t , the previous frame be I t-1 , the difference image J = |I t -I t-1 |, and set the threshold T. If the pixel value in the difference image is greater than T, it is considered that the position has changed:

[0100]

[0101] The result of the frame difference method can be directly used for foreground segmentation to quickly lock the position of the moving object. In combination with the gray scaled image, the human body posture can be more accurately located to provide key clues for subsequent behavior recognition.

[0102] S3: The feature extraction module adopts a multi-level neural network architecture for spatiotemporal graph construction, dynamic adjustment of graph structure, spatiotemporal attention weighting, and dynamic weight distribution, and outputs an optimized spatiotemporal feature matrix.

[0103] Further, the multi-level neural network architecture includes a spatio-temporal graph convolution network (ST-GCN) for spatio-temporal graph construction, an adaptive graph convolution network (AGCN) for dynamic adjustment of the graph structure, a spatio-temporal attention module (AST) for spatio-temporal attention weighting, and a skeletal joint enhancement traversal space configuration strategy for dynamic weight distribution.

[0104] The spatio-temporal graph convolution network converts the preprocessed image into skeletal joint information, constructs a spatio-temporal graph, extracts spatio-temporal features through multiple graph convolution layers, and constructs a feature matrix. Each layer of graph convolution transmits information between nodes through an adjacency matrix.

[0105] The input of the ST-GCN is preprocessed image data. First, the image is converted into skeletal joint information through a skeleton detection algorithm. This skeletal information includes the positions of various key points of the human body, forming a skeleton graph. The nodes of this graph represent different joints of the human body, and the edges represent the spatial relationships between the joints. Then, these skeleton information is used as a feature matrix X to construct a spatio-temporal graph. The nodes and edges in the graph change with the time sequence.

[0106] The ST-GCN extracts spatio-temporal features through graph convolution operations. Each layer of graph convolution transmits information between nodes through an adjacency matrix. This information transmission mechanism enables the network to effectively capture motion patterns and handle the combination of time series data and spatial structure. The formula for the graph convolution operation of the ST-GCN is as follows:

[0107]

[0108] Where: is the adjacency matrix with self-loops added, A is the original adjacency matrix representing the connection relationship between joints; I N is the identity matrix used to introduce self-loop connections; is the degree matrix, which represents the degree of nodes in the adjacency matrix; H (l) is the feature representation of the lth layer, W (l) is the weight matrix of the lth layer, and σ is the nonlinear activation function ReLU;

[0109] In the original description, the feature matrix X is the initial input of the graph convolution operation (i.e., H (0) = X), and is updated to Z (l+1) through multiple layers of graph convolution.

[0110] The feature matrix Z (l+1)The human body posture changes over time and space, and these spatio-temporal features will be input into the AGCN module; ST-GCN extracts spatio-temporal features from pre-processed bone joint information, which contains the change rule of human body posture and provides an important basis for subsequent behavior recognition.

[0111] Further, the adaptive graph convolutional network inputs a feature matrix output by the spatio-temporal graph convolutional network, which contains spatio-temporal features extracted by the spatio-temporal graph convolutional network.

[0112] ST-GCN (spatio-temporal graph convolutional network) is mainly used to extract features from spatio-temporal sequence data, and is particularly suitable for dynamic posture analysis based on skeletons. It can handle the combination of time series data and spatial structure information, helping the model understand the dynamic motion patterns obtained from time series images collected by drones. On this basis, AGCN (adaptive graph convolutional network) further optimizes the structure of graph convolution, enabling the model to adaptively adjust the edge weights of the graph structure, thereby better handling irregular data in complex environments (such as obstacles with different heights, angles, etc.), and improving recognition accuracy. Traditional GCN models have fixed topologies, limiting their flexibility.

[0113] The adaptive graph convolutional network introduces an adaptive adjacency matrix The graph convolutional network dynamically adjusts the edge weights and graph structure according to the actual relationship between data, so that AGCN can handle more complex scenarios and irregular data, thereby enhancing the model's ability to adapt to environmental changes;

[0114] The update method of the adaptive adjacency matrix is as follows:

[0115]

[0116] where Z (l+1) is the feature matrix of graph convolution in the spatio-temporal graph convolutional network, σ1 is the nonlinear activation function ReLU, Y (l +1) is the updated feature matrix, which dynamically adjusts the edge weights of the graph according to the learned parameters; W (l) is the weight matrix of the lth layer, is a new degree matrix calculated based on the adaptive adjacency matrix;

[0117] AGCN further optimizes the spatio-temporal features extracted from ST-GCN, making them better reflect the real changes of human body posture and improving the accuracy of behavior recognition.

[0118] Further, the spatio-temporal attention module aims to highlight important spatio-temporal regions and suppress irrelevant information to improve the sensitivity of the model to climbing behavior. The AST module is embedded in the feature extraction process, which can make the model focus on important action features by applying higher attention to key positions. The combination of AGCN and AST can enable the spatio-temporal graph convolutional network to not only adapt to environmental changes and adjust the structure of the graph when extracting spatio-temporal features, but also further strengthen the adaptive ability of the model through the attention mechanism. Specifically, AGCN processes the structure and spatial relationship of spatio-temporal data, while AST helps AGCN to focus on the most critical information in complex environments by weighting the importance of different time and space regions; in the AST module, we focus on key information by attention weighting these spatio-temporal features, thereby improving the recognition accuracy of climbing behavior.

[0119] Specifically, the updated feature matrix is attention weighted to calculate the spatial attention score ∈ i The importance of different joints is displayed, and the key joints are obtained according to the importance, highlighting the influence of the key joints on behavior recognition:

[0120]

[0121] where e i and e j are the non-normalized scores of nodes i and j, which are usually calculated by a lightweight neural network; ∈ i is the spatial attention score of node i, reflecting its importance in the current frame; N is the number of nodes, i, j ∈ N are variable indices, and is a positive constant.

[0122] The spatial attention score ∈ i is used to weight the updated feature matrix Y (l+1) to generate the spatial weighted feature matrix Y space :

[0123] Y space = ∈ i · Y (l+1)

[0124] where the attention mechanism is used to calculate the relevance between different time points. For each time step t, the non-normalized score of the current time step is calculated and normalized to the time attention score θ t :

[0125]

[0126] where f t and f kis the unnormalized score of time step t and time step k, usually computed by a lightweight neural network; β t is the time attention score of time step t, reflecting its importance in the time series; T is the total length of time steps, t, k ∈ T are variable indices, and β is a positive constant.

[0127] The time attention score θ t is used to weight the spatial weighted feature matrix Y space to generate the time weighted feature matrix Y time :

[0128] Y time = θ t ·Y space

[0129] where the time weighted feature matrix Y time is fused with the updated feature matrix Y (l+1) to generate the final weighted feature matrix of the spatio-temporal attention module

[0130]

[0131] where MLP is a multi-layer perceptron used to fuse attention weighted features; + represents a residual connection that preserves original feature information; Y time is the time weighted feature matrix;

[0132] Furthermore, the skeleton joint reinforced traversal space configuration strategy is used in the power pole tower climbing behavior monitoring system. The recognition of climbing behavior depends on the accurate analysis of human posture. However, in complex scenarios (such as changes in viewing angle, occlusion, and light interference), joint node positioning is often inaccurate, which affects the effect of behavior recognition. Traditional graph convolution networks rely on fixed adjacency matrices and are difficult to adapt to the dynamic relationship between joints in complex actions. Therefore, this strategy optimizes the spatial configuration between joints by introducing dynamic weight distribution, semantic distance measurement, and hierarchical traversal mechanism, enhancing the model's ability to analyze climbing behavior.

[0133] Specifically, the final weighted feature matrix Y output by the spatio-temporal attention module is set to have dimensions N × F, where N is the number of joints and F is the feature dimension.

[0134] The spatial configuration is optimized to decode the spatial coordinates (x i ,y i ,z i ) of the joints from the feature matrix, and the final weighted feature matrix Y is input. The importance weight w ijand generate the importance weight matrix:

[0135]

[0136] where || denotes feature concatenation, W w and b w are learnable parameters, and are the feature vectors of joint i and j respectively; the formula generates dynamic weights reflecting the semantic relationship between joints through feature concatenation and nonlinear transformation;

[0137] According to the importance weight, the weighted spatial distance is calculated and the weighted spatial distance matrix is generated. The traditional spatial distance formula only considers the Euclidean distance between joints (i.e. spatial distance). However, different joints have different importance in human posture and action. In order to enhance the influence of key joints, a weight factor can be introduced for each pair of joints, making the connection between joints more flexible and accurate. The spatial distance formula between joints is:

[0138]

[0139] where (x i ,y i ,z i ) and (x j ,y j ,z j ) are the spatial coordinates of joints i and j; w ij is the weight factor between joints i and j. This weight factor can be set according to the importance of the joint, such as the weight of the limb joint may be higher than that of the trunk joint, or in a certain action some joints have greater influence, according to the importance of different joints to obtain key joints, highlight the influence of key joints on behavior recognition.

[0140] Hierarchical traversal is performed on the weighted spatial distance, which is divided into local traversal and global traversal. The local traversal is to prioritize the connection between limb joints to capture local motion details; the global traversal aggregates local features based on dynamic weights to generate global motion pattern description;

[0141] In the hierarchical traversal process, the spatial relationship between joints is optimized through importance weight and weighted spatial distance, and the traversal result is combined with the final weighted feature matrix to generate the optimized joint feature matrix F ' , the formula is as follows:

[0142]

[0143] Wherein, WM is a dynamic weight matrix, Dis is a weighted spatial distance matrix, MLP is a multi-layer perceptron for fusing spatial optimization information.

[0144] In summary, the model can more finely analyze the changes in human posture, especially in complex actions such as climbing, and can more accurately identify abnormal behavior. Through this improved spatial configuration, the model can more finely analyze the dynamic changes of the human body, especially in complex behavior patterns (such as climbing behavior), thereby improving the recognition accuracy of these behaviors.

[0145] S4: The behavior recognition module adopts a CNN+Softmax architecture to extract local features through a convolutional neural network and calculate behavior probabilities through Softmax, outputting a climbing behavior probability value to determine whether there is a climbing behavior.

[0146] Specifically, the CNN+Softmax architecture includes a convolutional neural network and a Softmax behavior classification;

[0147] The convolutional neural network includes inputting the optimized joint feature matrix into the convolutional neural network of the behavior recognition module, including convolutional layers, pooling layers, and fully connected layers;

[0148] Local features of human posture are extracted through the convolutional layers:

[0149]

[0150] wherein U and V are the height and width of the convolution kernel, is the output value of the kth convolutional layer at joint nodes i and j, is the weight of the convolution kernel, is the input value of the k-1th layer at joint nodes i+u and j+v, and b k is the bias term.

[0151] The pooling layer subsamples the extracted local features, flattens the pooled features into a one-dimensional vector, and inputs them into the fully connected layer. The mathematical formula of max-pooling is as follows:

[0152]

[0153] wherein, is the output value of the kth pooling layer at positions i and j, s is the pooling step, and max represents the maximum value within the pooling window.

[0154] The fully connected layer maps the features processed by the pooling layer to the classification space to determine whether there is a climbing behavior. The mathematical formula of the fully connected layer is as follows:

[0155] z=W f ·x+bf

[0156] where z is the output of the fully connected layer, W f is the weight matrix, x is the flattened feature vector, and b f is the bias term.

[0157] The Softmax behavior classification includes calculating the probability distribution of each class using the Softmax function to determine whether it belongs to the climbing behavior:

[0158]

[0159] where P(y = i | x) is the probability that the input x belongs to class i, z i is the output value of the i-th neuron, i.e., the result of weighted summation; M is the total number of classes, and j is the variable index.

[0160] S5: The time series analysis module uses a long short-term memory network to perform time series analysis on the identified climbing behavior, and predicts the duration and trend of the climbing behavior.

[0161] Further, the long short-term memory network is used to perform time series analysis on the identified climbing behavior, and by constructing a long short-term memory network model, the duration and trend of the climbing behavior are predicted:

[0162] f t = σ2(W f · [h t-1 , x t ] + b f )

[0163] i t = σ2(W i · [h t-1 , x t ] + b i )

[0164]

[0165] o t = σ2(W o · [h t-1 , x t ] + b o )

[0166] h t = o t × tanh(C t )

[0167] where f t , i t , o toutput values of the forget gate, the input gate and the output gate, respectively, C t , t f i C o weight matrices of the forget gate, the input gate, the cell state and the output gate, respectively, b f i C o corresponding bias terms, respectively, σ2is a sigmoid activation function, and tanh is a hyperbolic tangent activation function.

[0168] S6: The early warning and response module implements a dynamic threshold strategy for behavior confirmation and trend verification, and triggers a response strategy to obtain a graded early warning signal and a response instruction, and implements global optimization through the system integration and optimization module.

[0169] It should be noted that behavior confirmation and trend verification include a detection stage, a time series analysis stage and an early warning stage.

[0170] The detection stage determines whether a climbing behavior occurs by comparing the probability value P(y=1|x) output by the behavior recognition module with a detection threshold T detect (such as T detect = 0.9) to trigger different levels of early warning signals.

[0171] When P(y=1|x) ≥ T detect , it is determined that a climbing behavior is confirmed, a red emergency early warning is performed, and an early warning signal is issued through sound and light alarms and platform pop-up windows, the alarm is automatically pushed to the monitoring center, a UAV or security personnel is started for on-site verification, and video evidence is recorded.

[0172] When P(y=1|x) < T detect but P(y=1|x) ≥ 0.5, it is determined that a suspected climbing behavior needs to be further verified in combination with the time series analysis stage.

[0173] The time series analysis stage predicts whether the climbing behavior will continue by analyzing the behavior probability trend of consecutive frames through the time series module, inputting the behavior probability sequence {P t , P t-1 ,..., P t-K} of the past K frames, and predicting the continuous probability of the behavior within Δt seconds in the future through the long short-term memory network of the time series module. If the continuous probability reaches a probability threshold P continue (such as P continue ≥ 0.8), it is determined that the climbing behavior will continue. ​​​​​​​​

[0174] When suspected climbing behavior and it is determined that the climbing behavior will continue, a yellow boundary warning is given and a warning signal is sent out through a platform pop-up window and a short message notification, manual review of the video stream is performed, and an emergency response is prepared;

[0175] When P(y=1∣x)<T detect , P(y=1∣x)<0.5 but there is abnormal motion, a blue observation warning is given and no warning signal is sent out through platform registration, the system automatically tracks the target, no active alarm is triggered, and only logs are recorded.

[0176] Embodiment 2, which is different from the previous embodiment, of the present application is as follows:

[0177] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0178] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a list of executable instructions for implementing logical functions, which can be specifically embodied in any computer readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions) or in conjunction with these instruction execution systems, devices or apparatus. For the purpose of this specification, "computer readable medium" can be any device that can contain, store, communicate, propagate or transport programs for use by an instruction execution system, device or apparatus or in conjunction with these instruction execution systems, devices or apparatus.

[0179] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection (electronic) having one or more wires, a portable computer diskette (magnetic), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer readable medium can be paper or another suitable medium upon which the program is printed, as the program can be electronically captured, for example via an optical scanner, then compiled, interpreted or otherwise processed in a suitable manner, if necessary, to generate an electronically readable version of the program, which can then be stored in the computer memory.

[0180] It should be understood that portions of the application can be implemented in hardware, software, firmware or combinations thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and as in another embodiment, can be implemented using any or a combination of the following technologies, which are well known in the art: a discrete logic circuit having logic gates for implementing logic functions upon data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0181] Embodiment 3, reference Figure 2 For the third embodiment of the present application, the embodiment provides an intelligent anti-climbing early warning system for power facilities, including an image acquisition module, a preprocessing module, a feature extraction module, a behavior recognition module, a time series analysis module, a warning and response module, and a system integration and optimization module.

[0182] The image acquisition module uses a fixed camera to capture image sequences of the tower area, as the basis for subsequent processing.

[0183] The preprocessing module uses the OpenCV library to preprocess the collected images according to the order of denoising, graying, morphological operation and frame difference method.

[0184] The feature extraction module is responsible for extracting key features for recognizing climbing behavior from the preprocessed images, including spatio-temporal graph convolution network, adaptive graph convolution network, spatio-temporal attention module and skeleton joint enhanced traversal space configuration strategy.

[0185] The behavior recognition module uses a trained neural network model to perform deep analysis on the features output by the feature extraction module to determine whether there is a climbing behavior.

[0186] a time series analysis module, which performs time series analysis and prediction on the identified climbing behavior by using a long short-term memory network in a time series analysis stage;

[0187] a warning and response module, which triggers a warning signal and formulates a corresponding response strategy when detecting the climbing behavior or predicting that the climbing behavior will continue to develop according to the prediction result of the time series analysis module;

[0188] a system integration and optimization module, which integrates all the modules to form a monitoring system and optimizes the algorithm.

[0189] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and all should be covered in the scope of the claims of the present application.

Claims

1. An intelligent anti-climbing early warning method for power facilities, characterized in that: The image acquisition module uses a fixed camera to capture image sequences of the utility pole tower area and outputs the raw image sequences to the preprocessing module. The preprocessing module uses OpenCV to perform preprocessing in the order of denoising, grayscale conversion, morphological operations, and frame difference method, and outputs the optimized grayscale image and motion region mask to the feature extraction module. The feature extraction module employs a multi-level neural network architecture to construct a spatiotemporal graph, dynamically adjust the graph structure, perform spatiotemporal attention weighting, and dynamically allocate weights, outputting an optimized spatiotemporal feature matrix. The behavior recognition module uses a CNN+Softmax architecture to extract local features through a convolutional neural network, calculate the behavior probability through Softmax, and output the climbing behavior probability value to determine whether climbing behavior exists. The time series analysis module uses a long short-term memory network to perform time series analysis on the identified climbing behavior and predict the duration and trend of the climbing behavior. The early warning and response module implements a dynamic threshold strategy to confirm behavior and verify trends, triggers response strategies, obtains tiered early warning signals and response instructions, and performs global optimization through the system integration and optimization module. The fixed camera is selected for its high resolution, wide dynamic range, night vision performance, and protective features; the original image sequence includes changes in lighting and moving targets.

2. The intelligent anti-climbing early warning method for power facilities as described in claim 1, characterized in that: The preprocessing includes using the OpenCV library to preprocess the acquired image in the following order: denoising, grayscale conversion, morphological operations, and frame difference method. The denoising process involves smoothing the video image and denoising the image using Gaussian filtering. The grayscale conversion transforms a color image into a grayscale image, retaining only the image's brightness information. The morphological operations remove noise and fill holes in the image by performing dilation and erosion operations. The dilation operation expands the white area in the binary image, and the erosion operation shrinks the white area. The morphological operations clean up irregular shapes on the image boundary. After denoising, grayscale conversion, and morphological operations, the image is processed using the frame difference method. The difference image is obtained by subtracting the previous frame from the current frame. Thresholding is then applied to highlight the changing areas. Combined with the grayscale image, the human pose is located.

3. The intelligent anti-climbing early warning method for power facilities as described in claim 2, characterized in that: The multi-level neural network architecture includes a spatiotemporal graph convolutional network for constructing a spatiotemporal graph and an adaptive graph convolutional network for dynamically adjusting the graph structure. The spatiotemporal graph convolutional network converts the preprocessed image into skeletal joint information, constructs a spatiotemporal graph, extracts spatiotemporal features through multiple graph convolutional layers, and constructs a feature matrix. Each graph convolutional layer transmits information between nodes through an adjacency matrix. The adaptive graph convolutional network takes as input the feature matrix output by the spatiotemporal graph convolutional network, which contains the spatiotemporal features extracted by the spatiotemporal graph convolutional network. Adaptive graph convolutional networks introduce an adaptive adjacency matrix. The graph convolutional network dynamically adjusts edge weights and graph structure based on the relationships between actual data. The adaptive adjacency matrix update method is as follows: Among them, Z (l+1) Let σ1 be the feature matrix of graph convolution in the spatiotemporal graph convolutional network, and let σ1 be the nonlinear activation function ReLU. It is an adjacency matrix with self-loops added, where A is the original adjacency matrix, representing the connection relationships between key nodes; I N Y is the identity matrix used to introduce self-loop connections. (l+1) It is the updated feature matrix, and the edge weights of the graph are dynamically adjusted based on the learned parameters; W (l) It is the weight matrix of the l-th layer. It is a new degree matrix calculated based on the adaptive adjacency matrix.

4. The intelligent anti-climbing early warning method for power facilities as described in claim 3, characterized in that: The multi-level neural network architecture also includes a spatiotemporal attention module for spatiotemporal attention weighting and a skeletal joint enhanced traversal space configuration strategy for dynamic weight allocation. The spatiotemporal attention module performs attention weighting on the updated feature matrix, calculates spatial attention scores to display the importance of different key points, and obtains key key points based on the importance results of different key points, highlighting the impact of key key points on behavior recognition. The updated feature matrix is ​​weighted using spatial attention scores to generate a spatially weighted feature matrix. The attention mechanism calculates the correlation between different time steps. For each time step, the unnormalized score of the current time step is calculated and normalized to the time attention score by the Softmax function. The spatially weighted feature matrix is ​​weighted using the temporal attention score to generate a temporally weighted feature matrix; The time-weighted feature matrix is ​​fused with the updated feature matrix to generate the final weighted feature matrix for the spatiotemporal attention module. Wherein, MLP stands for Multilayer Perceptron, used to fuse attention-weighted features; + indicates residual connection, preserving original feature information; Y time The time-weighted feature matrix; The skeletal joint enhancement traversal space configuration strategy will enhance the final weighted feature matrix output by the spatiotemporal attention module. Define the dimension as N×F, where N is the number of key points and F is the feature dimension; Spatial configuration optimization is performed by decoding the spatial coordinates of the joints from the feature matrix, inputting the final weighted feature matrix, calculating the importance weights of the dynamic weights between each pair of joints through a lightweight network and generating an importance weight matrix, and calculating the weighted spatial distance based on the importance weights and generating a weighted spatial distance matrix. The weighted spatial distance is traversed hierarchically, which is divided into local traversal and global traversal. The local traversal prioritizes the traversal of the connections between the joints of the limbs to capture local motion details. The global traversal aggregates local features based on dynamic weights to generate a global motion pattern description. During the hierarchical traversal, the spatial relationships between key points are optimized through importance weights and weighted spatial distances. The traversal results are then compared with the final weighted feature matrix. Combined, an optimized key feature matrix F is generated. ' The formula is as follows: Wherein, WM is the dynamic weight matrix, Dis is the weighted spatial distance matrix, and MLP is the multilayer perceptron, used to fuse spatial optimization information.

5. The intelligent anti-climbing early warning method for power facilities as described in claim 4, characterized in that: The determination of whether climbing behavior exists includes using a CNN+Softmax architecture that includes a convolutional neural network and Softmax behavior classification. The convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, into which the optimized key feature matrix is ​​input into the behavior recognition module. Local features of human pose are extracted using convolutional layers; The pooling layer downsamples the extracted local features, flattens the pooled features into a one-dimensional vector, and inputs it into the fully connected layer. The fully connected layer maps the features processed by the pooling layer to the classification space to determine whether climbing behavior exists; The Softmax behavior classification includes calculating the probability distribution of each category using the Softmax function to determine whether it belongs to climbing behavior: Where P(y=i|x) is the probability that input x belongs to class i, and z i is the output value of the i-th neuron, i.e., the result after weighted summation; M is the total number of categories, and j is the variable index.

6. The intelligent anti-climbing early warning method for power facilities as described in claim 5, characterized in that: The time series analysis includes using a long short-term memory network to perform time series analysis on the identified climbing behavior, and predicting the duration and trend of the climbing behavior by constructing a long short-term memory network model: f t =σ2(W f ·[h t-1 ,x t ]+b f ) i t =σ2(W i ·[h t-1 ,x t ]+b i ) o t =σ2(W o ·[h t-1 ,x t ]+b o ) h t =o t ×tanh(C t ) Among them, f t i t o t The output values ​​of the forget gate, input gate, and output gate are respectively, C. t , These represent the cell state and the candidate cell state, respectively. t W is the output value of the hidden layer. f W i W C W o Here are the weight matrices for the forget gate, input gate, cell state, and output gate, respectively. f b i b C b o These are the corresponding bias terms, where σ2 is the sigmoid activation function and tanh is the hyperbolic tangent activation function.

7. The intelligent anti-climbing early warning method for power facilities as described in claim 6, characterized in that: The behavior confirmation and trend verification include a detection phase, a time series analysis phase, and an early warning phase; The detection phase determines whether climbing behavior has occurred, based on the probability value P (y = 1 | x) output by the behavior recognition module and the detection threshold T. detect By comparing data, different levels of warning signals can be triggered: When P(y=1|x)≥T detect If the climbing behavior is confirmed, a red emergency warning will be issued and a warning signal will be sent through sound and light alarms and platform pop-ups. The alarm will be automatically pushed to the monitoring center, and drones or security personnel will be activated for on-site verification, while video evidence will be recorded. When P(y=1|x) <T detect However, P(y=1|x)≥0.5 is considered a suspected climbing behavior, which needs to be further verified in conjunction with the time series analysis stage; The time series analysis phase predicts whether climbing behavior will continue. It analyzes the probability trend of behavior across consecutive frames using a time series module, inputting the probability sequence of behavior from the past K frames {P}. t ,P t-1 ,...,P t-K The long short-term memory network of the time series module predicts the probability of the behavior lasting within Δt seconds. If the probability of the behavior lasting within Δt seconds reaches the probability threshold P, then... continue If so, the climbing behavior will continue; When suspected climbing behavior is detected and it is determined that the climbing behavior will continue, a yellow warning will be issued and a warning signal will be sent through platform pop-ups and SMS notifications. The video stream will be manually reviewed and an emergency response will be prepared. When P(y=1|x) <T detect If P(y=1|x)<0.5 but abnormal movement exists, a blue observation warning is issued and the abnormality is registered through the platform without sending a warning signal. The system automatically tracks the target, does not trigger an active alarm, and only records the log.

8. An intelligent anti-climbing early warning system for power facilities, employing the intelligent anti-climbing early warning method for power facilities as described in any one of claims 1 to 7, characterized in that: It includes an image acquisition module, a preprocessing module, a feature extraction module, a behavior recognition module, a time series analysis module, an early warning and response module, and a system integration and optimization module; The image acquisition module uses a fixed camera to capture image sequences of the utility pole tower area as the basis for subsequent processing; The preprocessing module uses the OpenCV library to preprocess the acquired image in the following order: denoising, grayscale conversion, morphological operations, and frame difference method. The feature extraction module is responsible for extracting key features for recognizing climbing behavior from the preprocessed image, including a spatiotemporal graph convolutional network, an adaptive graph convolutional network, a spatiotemporal attention module, and a skeletal joint enhanced traversal space configuration strategy. The behavior recognition module uses a trained neural network model to perform in-depth analysis on the features output by the feature extraction module to determine whether climbing behavior exists. The time series analysis module, in the time series analysis stage, uses a long short-term memory network to perform time series analysis and prediction on the identified climbing behavior; The early warning and response module, based on the prediction results of the time series analysis module, triggers an early warning signal and formulates a corresponding response strategy when it detects climbing behavior or predicts that the climbing behavior will continue to develop. The system integration and optimization module integrates all modules to form a monitoring system, and at the same time optimizes the algorithm.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent anti-climbing early warning method for power facilities as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent anti-climbing early warning method for power facilities as described in any one of claims 1 to 7.

Images